Rotating equipment fault diagnosis method fusing cnn and graph attention network

Abstract

The application provides a rotating equipment fault diagnosis method combining CNN and a graph attention network. Multiple sensors capture key operation signals of each part of a motor, a bearing, a gear box, a shaft coupling and a rotating load, and the operation signals are input into a multi-channel attention module through a GRU network to output a feature sequence. The feature sequence is spliced into a two-dimensional feature matrix, which is input into a CNN and a hybrid attention mechanism module to obtain fault features of each part. The fault features are windowed according to a time sequence, and are reorganized according to dynamic weighting based on importance. The reorganized features of different parts are used as nodes, and significant interaction relationships between the nodes are used as edges to construct a global fault diagnosis graph. The interaction information between the nodes is learned and aggregated through a graph attention network to obtain global fault features. A graph network is constructed to optimize a loss function, and model parameters are analyzed and optimized. The model fully excavates fault information in a multi-stage fusion mode, and greatly improves the accuracy of fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of hardware device fault diagnosis driven by big data deep network models, specifically involving a fault diagnosis method for rotating devices that integrates CNN and graph attention network. Background Technology

[0002] Rotating equipment, as the "heart" of industrial production systems, is widely used in critical fields such as aerospace, energy and power, and petrochemicals. Its stable operation is a core element in ensuring production continuity, safety, and economy. However, rotating equipment operates under long-term high loads and complex conditions, making it highly susceptible to various fault types, such as bearing wear, gear failure, and rotor imbalance. Once a fault occurs, it not only leads to equipment downtime and production halts but can also trigger a chain reaction, causing incalculable economic losses and safety hazards. The fault characteristics of rotating equipment are complex and varied; fault signals often exhibit nonlinear and non-stationary characteristics, and different fault modes are coupled, significantly increasing the difficulty of fault diagnosis. Simultaneously, rotating equipment generates multi-source heterogeneous data such as vibration, temperature, and current during operation, with complex spatial correlations and dynamic dependencies among these data. The rotating equipment fault diagnosis method proposed in this application, which integrates CNN and graph attention networks, can comprehensively and deeply mine fault characteristics, achieving accurate fault diagnosis for rotating equipment.

[0003] Currently, traditional fault diagnosis methods, such as those based on vibration spectrum analysis and threshold judgment, rely excessively on human experience, resulting in low diagnostic efficiency and accuracy, and failing to meet the diagnostic needs under complex operating conditions. While convolutional neural networks (CNNs) in deep learning can automatically extract local features from fault data, they have limitations when dealing with the complex spatial relationships between multi-source data from rotating equipment. Graph attention networks, although effective in capturing the correlation information between data nodes, struggle to deeply mine the local detailed features of the data. Therefore, there is an urgent need for a fault diagnosis method that integrates CNNs and graph attention networks, fully leveraging their respective advantages to achieve rapid and accurate identification of faults in rotating equipment. This is of significant practical importance for reducing equipment operation and maintenance costs and improving the level of intelligent industrial production. Summary of the Invention

[0004] The purpose of this invention is to provide a more accurate and reliable method for diagnosing faults in rotating equipment, which can accurately detect faults during the operation of rotating equipment and ensure the safe and stable operation of the equipment.

[0005] To address the aforementioned technical problems, this invention proposes a fault diagnosis method for rotating devices that integrates CNN and graph attention networks, comprising:

[0006] Multiple sensors are used to collect key operating signals from different parts: ① current, voltage, temperature and vibration signals are collected from the motor; ② vibration and temperature signals are collected from the bearing; ③ vibration and temperature signals are collected from the gearbox; ④ vibration, temperature and sound signals are collected from the coupling; ⑤ speed, temperature and vibration signals are collected from the rotating load. Each signal is input into its respective GRU network, and the hidden layer state of each time step of the GRU network is used as the input of one channel of the channel attention module. The multi-channel attention module generates feature vectors by weighting different channels. The feature vectors generated by each channel are concatenated along the channel dimension so that each row of the two-dimensional feature matrix corresponds to the feature vector of one channel, and each column corresponds to a feature point in the feature vector. The two-dimensional feature matrix is ​​input into the CNN and the hybrid attention module to extract features and their weights at a deep level. The four-layer CNN network uses 7*7, 3*3, 3*3 and 1*1 multi-scale convolution kernels respectively, and a hybrid attention mechanism is connected after each layer of CNN for weighting and local adjustment. Finally, refined fault features of different signals from different parts of the rotating equipment are obtained.

[0007] The refined fault characteristics of different signals from various parts of the rotating equipment are divided into N time windows according to their time characteristics at fixed time intervals, as shown in formula (1), where T is the total number of time steps, W is the window size, and S is the step length. For the i-th time window, the similarity between two features at different time steps is calculated according to formula (2). This allows us to obtain the similarity matrix for the i-th time window. Where j and k are two different time steps, and f is the feature of the time step, the weight matrix of the time window is obtained after normalization using the softmax function. The features within the time window are weighted according to formula (3) based on the weight matrix. Local aggregated features are obtained after weighting. ,in Let the feature of the i-th window at time j be the feature of all windows. By splicing the data in chronological order, the time window splicing characteristics of different signals from each part are finally obtained. , where N is the number of windows and d is the feature dimension;

[0008] 1(1)

[0009] (2)

[0010] (3)

[0011] The electric drive system is viewed as a graph topology, with the motor, bearings, gearbox, coupling, and rotating load considered as nodes of the graph network, and the connections or coupling relationships between them as edges. The time window splicing characteristics of different signals from each part are used as the node feature information of the graph network. The system includes: ① Motor nodes, which include four signal features: current harmonic distortion rate, stator mode frequency response, voltage fluctuation, and temperature; ② Bearing nodes, which include two signal features: inner and outer ring fault characteristic frequencies and temperature; ③ Gearbox nodes, which include two signal features: gear meshing frequency and temperature; ④ Coupling nodes, which include three signal features: vibration, temperature, and sound; and ⑤ Rotating load nodes, which include three signal features: speed, temperature, and vibration. The degree of fault propagation between two nodes is used as edge information to construct a global fault diagnosis graph.

[0012] The features of each node are mapped to Query and Key value vectors through a learnable linear transformation matrix. The shared attention mechanism function LeakyReLU is used to calculate the attention coefficients between different nodes, and the attention coefficients are modulated by combining edge information. Finally, all attention coefficients are normalized to obtain the attention weights. The node features are weighted, and the weighted features of different nodes are learned in different subspaces through a multi-head attention mechanism. Finally, the outputs of each head are integrated to obtain the global fault feature Z.

[0013] Constructing a graph network to jointly optimize the loss function Analyze and adjust the model parameters.

[0014] Optionally, the constructed graph network jointly optimizes the loss function. Analyze and adjust model parameters, including:

[0015] ① Construct a graph network joint optimization loss function As shown in equation (4):

[0016] (4)

[0017] in For classification loss function, ② Calculate the global-local structural contrast loss function. As shown in formula (5):

[0018] (5)

[0019] Where Z represents the global fault characteristic. Characteristics of node failures. These are the weighting coefficients. It reflects the characteristics of global faults and node faults. ③ Calculate the classification loss function based on the difference between the means. As shown in formula (6):

[0020] (6)

[0021] Where C represents the total fault category. For the sample Encoding in dataset c To predict the probability that the fault is of type c for the model, ④ based on the joint optimization loss function. Update model parameters, including: update the linear transformation matrix. and scaling factor ,by For example, as shown in formula (7):

[0022] (7)

[0023] Where t is the number of model iterations, This is the learning rate.

[0024] This application proposes a fault diagnosis method for rotating equipment that integrates CNN and graph attention network. Multiple sensors capture key operating signals from various components, including the motor, bearings, gearbox, coupling, and rotating load. These signals are input into a multi-channel attention module via a GRU network, and the output is a feature sequence. This feature sequence is concatenated into a two-dimensional feature matrix, which is then input into a CNN and hybrid attention mechanism module to obtain fault features for each component. The fault features are then windowed according to time series and dynamically weighted based on importance before being recombined. A global fault diagnosis graph is constructed using the recombined features from different components as nodes and the significant interactions between nodes as edges. The graph attention network learns and aggregates the interaction information between nodes to obtain the global fault features. Finally, a graph network is constructed to jointly optimize the loss function. The model parameters were analyzed and optimized. This model, through multi-stage fusion, comprehensively mines fault information, greatly improving the accuracy of fault diagnosis. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic flowchart of a rotating device fault diagnosis method that integrates CNN and graph attention network provided in an embodiment of this application.

[0027] Figure 2 A model structure diagram provided for an embodiment of the present invention;

[0028] Figure 3 A graph attention algorithm framework diagram provided for embodiments of the present invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] like Figure 1 As shown, Figure 1 The flowchart of the rotating device fault diagnosis method that integrates CNN and graph attention network provided in the embodiments of this application is shown, and specifically includes 5 parts.

[0031] S11: Data is collected from multiple locations, and after being weighted by a multi-channel attention mechanism through a GRU network and extracted by a CNN and a hybrid attention mechanism, refined fault features of different signals for each location are obtained.

[0032] It should be noted that the raw data collected from different locations have varying data scales and feature importance. Directly extracting fault features can lead to biased results. Therefore, the raw data must be normalized and the data scale standardized before feature extraction. Secondly, the different equipment operation data collected from each location are input into a multi-channel attention module through a GRU network, with the hidden layer state of the GRU network at each time step serving as the input to one channel. Finally, the weighted concatenation of features from each channel is processed by a CNN and a hybrid attention module to mine deeper information from the data, thereby obtaining refined fault features for different signals in each location. The entire process mainly includes three steps.

[0033] Step 11: Normalize the collected raw data using formula (1): (1)

[0034] Where t represents the time step and i represents the location. Let i be the maximum and minimum values ​​of all time points in part i, respectively. The normalized data is input into the GRU network. Important data features are extracted and retained from historical data through reset and update gates. The update gate state at the current time step is calculated using the input of the current state and the output of the hidden layer of the previous state. and reset door status As shown in formulas (2) and (3):

[0035] (2)

[0036] (3)

[0037] Based on the current state and The output of the candidate hidden state at this moment is calculated as shown in formula (4):

[0038] (4)

[0039] Finally, combining the hidden state from the previous time step... and candidate hidden state Calculate the hidden state at the current time step, as shown in formula (5):

[0040] (5)

[0041] In the above formula The weight matrix representing model training. , , Represents the bias matrix. and tanh represent the sigmoid and tanh activation functions, respectively.

[0042] After obtaining the hidden layer states of each GRU network at each time step Then, a multi-channel attention module is used to weight the data to highlight the importance of data at different times, for each... Calculate its attention score using formula (6). Attention weights are obtained by normalizing all attention scores using the softmax function. As shown in formula (7), the attention weights are finally compared with the corresponding hidden layer states. The weights are calculated using formula (8) to obtain the weighted feature vectors for each channel. .

[0043] (6)

[0044] (7)

[0045] (8)

[0046] Step 12: Concatenate the feature vectors generated by each channel along the channel dimension so that each row of the two-dimensional feature matrix corresponds to the feature vector of one channel, and each column corresponds to a feature point in the feature vector.

[0047] Step 13: Input the concatenated two-dimensional feature matrix into the module that appears 4 times in the CNN and the hybrid attention mechanism. The first layer of the CNN network uses a 7*7 convolution kernel to perform convolution operation to obtain the feature map F. Then, the activation function Swish(x) = xsigmoid(x) is used to introduce nonlinearity and enhance the expressive power of the model. After the convolutional layer, the hybrid attention module is connected. The feature map is weighted and locally adjusted by parallel channel attention and spatial attention, as follows. The channel attention weights are calculated according to formulas (9), (10) and (11). Spatial attention weights are calculated according to formula (12). :

[0048] (9)

[0049] (10)

[0050] (11)

[0051] (12)

[0052] in, , Here is the weight matrix, and c is the channel label.

[0053] The formula for calculating the feature map F1' is shown in equation (13):

[0054] (13)

[0055] In the subsequent second, third, and fourth iterations of the CNN and hybrid attention mechanism modules, the above steps were repeated using 3x3, 3x3, and 1x1 convolutional kernels, respectively, to obtain refined fault feature sequences for different signals in each part. , where T=1000.

[0056] Based on the above theory, in one optional embodiment of this application, a 4-layer CNN and a hybrid attention module are arranged sequentially. From front to back, each CNN layer uses convolutional kernels of sizes 7*7, 3*3, 3*3, and 1*1, respectively, ultimately obtaining a refined fault feature sequence with 1000 time steps. .

[0057] S12: Window the fault features according to the time series and reorganize them after weighting based on dynamics.

[0058] It should be noted that CNN and hybrid attention modules do not focus on the temporal characteristics of the data, but the regular movement of the convolutional kernels preserves the temporal order of the features to a certain extent. After windowing and weighting, it is easier to capture various types of information about the features in a comprehensive manner. The whole process mainly involves three steps.

[0059] Step 21: For the obtained refined fault feature sequence According to equation (14), it is divided into N time windows of size W and step size S=50, where the i-th time window can be represented as .

[0060] 1 (14)

[0061] Step 22: For each time window The similarity between features at two different time steps is calculated using equation (15). This allows us to obtain the similarity matrix for the i-th time window. Where j and k are two different time steps, and f is the time step feature, the similarity matrix is... Each element in the matrix is ​​normalized using the softmax function as shown in equation (16), thereby obtaining the weight matrices for different time windows. And calculate the mean of the corresponding weights according to formula (17) to obtain the weight at each time step. , where i is the time window number and j is the time step.

[0062] (15)

[0063] (16)

[0064] (17)

[0065] Step 23: Apply the corresponding weight matrix to the features within each time window. Local features are obtained by weighting according to formula (18). and all time windows By stitching the pieces together sequentially in chronological order, the final time window stitching feature is obtained. .

[0066] (18)

[0067] Based on the above theory, in one optional embodiment of this application, each refined fault feature is divided into 20 time windows, and the weighted summation and recombination yields the time window splicing feature. .

[0068] S13: Construct a global fault diagnosis graph using the weighted and recombined fault features as nodes and the significant interaction relationships between nodes as edges.

[0069] It should be noted that there are direct or indirect connections or couplings between the motor, bearings, gearbox, coupling and rotating load in the point drive system. Constructing them into a graph network is beneficial for analyzing the relationships between the signal characteristics of each part and facilitating the global analysis of fault characteristics. The whole process mainly involves two steps.

[0070] Step 31: Treat the electric drive system as a graph topology, with the motor, bearings, gearbox, coupling, and rotating load as nodes of the graph network, and the connections or coupling relationships between them as edges. Use the time window splicing features of different signals from each part as the node feature information of the graph network. Among them: ① four signal characteristics, namely current harmonic distortion rate, stator mode frequency response, voltage fluctuation and temperature, are used as motor node characteristic information; ② two signal characteristics, namely inner and outer ring fault characteristic frequency and temperature, are used as bearing node characteristic information; ③ two signal characteristics, namely gear meshing frequency and temperature, are used as gearbox node characteristic information; ④ three signal characteristics, namely vibration, temperature and sound, are used as coupling node characteristic information; ⑤ three signal characteristics, namely speed, temperature and vibration, are used as rotating load node characteristic information.

[0071] Step 32: Based on the equipment's historical operating data and the system's operating conditions, define the degree of fault propagation between each pair of the five nodes: motor, bearing, gearbox, coupling, and rotating load. This serves as the edge information of the graph network.

[0072] Based on the above theory, in one optional embodiment of this application, the number of nodes in the graph is 5. The motor node includes four signal characteristics: current harmonic distortion rate, stator modal frequency response, voltage fluctuation, and temperature. The bearing node includes two signal characteristics: inner and outer ring fault characteristic frequencies and temperature. The gearbox node includes two signal characteristics: gear meshing frequency and temperature. The coupling node includes three signal characteristics: vibration, temperature, and sound. The rotating load node includes three signal characteristics: speed, temperature, and vibration. The degree of fault propagation influence between two nodes with a direct physical connection is also considered. =0.7, the degree of impact of fault propagation between two nodes that do not have a direct physical connection. =0.15.

[0073] S14: Learn and aggregate fault information between different nodes through a graph attention network to obtain global fault features.

[0074] It should be noted that the weighted fault features of each node may have one or more correlations. This application selects a multi-head attention mechanism to capture multiple relationships between the features of different nodes. The whole process mainly includes three steps.

[0075] Step 41: Obtain the node feature set V from the global fault diagnosis map as the input matrix. Where N is the number of nodes, D is the feature dimension of each node, and a shared linear transformation matrix is ​​used. The node features are mapped to Query and Key vectors according to equations (19) and (20):

[0076] (19)

[0077] (20)

[0078] And as shown in formula (21):

[0079] (twenty one)

[0080] Calculate the similarity between nodes as the original attention coefficient. Where i and j represent two distinct nodes, This is a scaling factor to prevent the dot product result from becoming too large and causing the gradient to vanish.

[0081] Step 42: Use the degree of fault propagation between the two nodes as the original attention coefficient. The modulation factor is obtained, and the modulation attention coefficient is calculated according to formula (22). .

[0082] (twenty two)

[0083] Step 43: Use the LeakyRelu function to perform a nonlinear transformation on the original coefficients to alleviate gradient vanishing, as shown in equation (23), where Weighting coefficients:

[0084] (twenty three)

[0085] The original attention coefficients are then normalized to transform them into attention weights in the form of a probability distribution, resulting in... , representing the normalized attention weight of node i to node j, ultimately yielding the attention weight matrix. .

[0086] Step 44: Use the attention weight matrix B to weight the node features to obtain the weighted features. And the weighted levy Processing is performed in parallel using h independent heads, repeating steps 41 and 42, where each head corresponds to an independent linear transformation matrix. This process yields independent Query, Key, and Value values ​​for each head, leading to different weight matrices. Different heads learn different fault information and relationships between nodes. Finally, the outputs of each head are concatenated and dimensionality is reduced through linear transformation to obtain the final global fault feature Z.

[0087] Based on the above discussion, in one optional embodiment of this application, the multi-head attention mechanism has a total of 6 heads, and the coefficients in the LeakyReLU function... The value is 0.01.

[0088] S15: Construct a graph network to jointly optimize the loss function, analyze and adjust the model parameters.

[0089] It should be noted that the global fault features learned by the graph attention network may differ significantly from the fault features of individual nodes, and there may also be some errors in fault feature classification. Therefore, constructing a graph network to jointly optimize the loss function can calculate the loss magnitude and optimize the model parameters accordingly, thereby improving the model's fault diagnosis and classification performance. The entire process mainly involves two steps:

[0090] Step 51: Construct the graph network joint optimization loss function As shown in formula (24):

[0091] (twenty four)

[0092] in, The classification loss function is extended from the binary cross-entropy to the multi-class scenario. It can measure the difference between the predicted probability distribution and the true distribution. The calculation formula is shown in Equation (25):

[0093] (25)

[0094] Where C represents the total fault category. For the sample Encoding in dataset c Predict the probability that the fault belongs to category c for the model; The global-local structural loss function can calculate whether the mean values ​​of the global fault representation and the local node features are consistent, thus avoiding the loss of fine-grained fault information during feature aggregation. The calculation formula is shown in equation (26):

[0095] (26)

[0096] Where Z represents the global fault characteristic. Characteristics of node failures. These are the weighting coefficients.

[0097] Step 52: Calculate the joint optimization loss function of the graph network. First, calculate according to formulas (27) and (28). For linear transformation matrix (by (For example) and scaling factor (Initial value) The gradient of )

[0098] (27)

[0099] (28)

[0100] in The weights for the classification loss are determined according to the joint optimization loss function. The model parameters are updated according to the gradient of each parameter using formula (29). , (with linear transformation matrix) (For example).

[0101] (29)

[0102] Based on the above theory, the global-local structural loss function is calculated. middle The value is generally between 0.5 and 1.0. (Calculation) For linear transformation matrix gradient in The value is 0.05.

[0103] This application uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are merely for the purpose of helping to explain the method and core ideas of the present invention. It should be noted that those skilled in the art can make various improvements and modifications to the present invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A method for diagnosing faults in rotating equipment by integrating CNN and graph attention network, characterized in that, include: Multiple sensors are used to collect key operating signals from different parts: ① Current, voltage, temperature and vibration signals are collected from the motor; ② Vibration and temperature signals are collected from the bearing; ③ Vibration and temperature signals are collected from the gearbox; ④ Vibration, temperature and sound signals are collected from the coupling; ⑤ Rotational speed, temperature and vibration signals are collected from the rotating load. Each signal is input into its respective GRU network, and the hidden layer state of each time step of the GRU network is used as the input of one channel of the multi-channel attention module. The multi-channel attention module generates feature vectors by weighting different channels. The feature vectors generated by each channel are concatenated along the channel dimension so that each row of the two-dimensional feature matrix corresponds to the feature vector of one channel, and each column corresponds to a feature point in the feature vector. The two-dimensional feature matrix is ​​input into the CNN and the hybrid attention module to extract features and their weights at a deep level. The four-layer CNN network uses 7*7, 3*3, 3*3 and 1*1 multi-scale convolutional kernels respectively, and a hybrid attention mechanism is connected after each CNN layer for weighting and local adjustment. Finally, refined fault features of different signals from different parts of the rotating equipment are obtained. The refined fault characteristics of different signals from various parts of the rotating equipment are divided into N time windows according to their time characteristics at fixed time intervals, as shown in formula (1), where T is the total number of time steps, W is the size of the time window, and S is the step size. For the i-th time window, the similarity between the features of two different time steps is calculated according to formula (2). This allows us to obtain the similarity matrix for the i-th time window. Where j and k are two different time steps, and f is the feature of the time step, the weight matrix of the time window is obtained after normalization using the softmax function. The features within the time window are weighted according to formula (3) based on the weight matrix. Local aggregated features are obtained after weighting. ,in For the feature of the j-th time step in the i-th time window, and combine the features of all time windows... By splicing the data in chronological order, the time window splicing characteristics of different signals from each part are finally obtained. , where N is the number of time windows and d is the feature dimension; 1（1） （2） （3） The electric drive system is viewed as a graph topology, with the motor, bearings, gearbox, coupling, and rotating load within it considered as nodes of the graph network, and the connections or coupling relationships between nodes as edges. The time window splicing characteristics of different signals from each part are used as the node feature information of the graph network. Specifically: ① Motor nodes include four signal features: current harmonic distortion rate, stator mode frequency response, voltage fluctuation, and temperature; ② Bearing nodes include two signal features: inner and outer ring fault characteristic frequencies and temperature; ③ Gearbox nodes include two signal features: gear meshing frequency and temperature; ④ Coupling nodes include three signal features: vibration, temperature, and sound; ⑤ Rotating load nodes include three signal features: speed, temperature, and vibration. The degree of fault propagation influence between two nodes is used as edge information to construct a global fault diagnosis graph. The features of each node are mapped to Query and Key value vectors through a learnable linear transformation matrix. The shared attention mechanism function LeakyReLU is used to calculate the attention coefficients between different nodes, and the attention coefficients are modulated by combining edge information. Finally, all attention coefficients are normalized to obtain the attention weights. The node features are weighted, and the weighted features of different nodes are learned in different subspaces through a multi-head attention mechanism. Finally, the outputs of each head are integrated to obtain the global fault feature Z. Constructing a graph network to jointly optimize the loss function Analyze and adjust the model parameters.

2. The method for diagnosing faults in rotating equipment by fusing CNN and graph attention networks as described in claim 1, characterized in that, Constructing a graph network to jointly optimize the loss function Analyze and adjust model parameters, including: ① Construct a graph network joint optimization loss function As shown in equation (4): （4） in For classification loss function, ② Calculate the global-local structural contrast loss function. As shown in formula (5): （5） Where Z represents the global fault characteristic. Characteristics of node failures. These are the weighting coefficients. It reflects the characteristics of global faults and node faults. ③ Calculate the classification loss function based on the difference between the means. As shown in formula (6): （6） Where C represents the total fault category. The encoding of sample m in dataset c. To predict the probability that the fault is of type c for the model, ④ based on the joint optimization loss function. Update model parameters, including: update the linear transformation matrix. and scaling factor As shown in formula (7): （7） Where t is the number of model iterations, For the learning rate, and and scaling factor The calculation formula and t in the formula, , All with The calculation formula is the same.

Images